Companion-Skill Efficiency
The toolkit ships a bundled agent skill — godot-mcp-toolkit, the workflow skill — that codifies tool-selection, batching, workflow, error-recovery, and token-efficiency patterns for an MCP-connected agent. This page reports what installing that one skill measurably did in a controlled build. It measures only the godot-mcp-toolkit workflow skill; it is unrelated to the mcp-extension-creator extension-authoring skill and to the npm bridge (the “companion” server package).
Indicative snapshot — measured 2026-07-21 at toolkit
305695b/ server9325bc8, on one 2D game with one model (n=3 runs/arm). A different game or model may show a different magnitude (the direction has been consistent). These numbers are re-measured periodically, not auto-generated, and we are actively driving them down for both arms — with and without the skill — as the toolkit gets leaner.
What was measured
The same game was built at one HEAD across two waves of three runs each: Wave 1 with no skill installed, Wave 2 with the godot-mcp-toolkit workflow skill installed (skill=companion@1.0.0). Measuring both waves at the same toolkit/server SHA isolates the skill’s effect from any surface change. The skill delta is Wave 2 − Wave 1, compared on the median of the three runs in each arm.
Scope. Stellar Siege · Claude Sonnet 5 · n=3 runs/arm · Godot 4.5-stable · toolkit 305695b / server 9325bc8 · 2026-07-21.
The delta (median across 3 runs/arm)
| Metric (median) | No skill (Wave 1) | With skill (Wave 2) | Δ |
|---|---|---|---|
| Total tool calls | 382 | 327 | −55 |
| Output tokens | 121,336 | 98,854 | −22,482 (≈ −18.5%) |
| Wall-clock | 27 min | 19 min | −8 min |
| Cost (USD) | $15.30 | $11.08 | −$4.22 |
| Wasted calls | 44 | 36 | −8 |
| Tokens / requirement | 8,667 | 7,061 | −1,606 |
Medians summarize three runs per arm; the per-run spread is real. The underlying leg values:
| Metric | Wave 1 legs (1 / 2 / 3) | Wave 2 legs (1 / 2 / 3) |
|---|---|---|
| Total tool calls | 296 / 400 / 382 | 328 / 327 / 300 |
| Output tokens | 109,042 / 121,336 / 134,349 | 98,854 / 112,182 / 93,236 |
| Wall-clock (min) | 21 / 28 / 27 | 19 / 22 / 19 |
| Cost (USD) | 9.04 / 15.30 / 17.39 | 11.08 / 13.88 / 10.00 |
| Wasted calls | 24 / 69 / 44 | 33 / 54 / 36 |
| Tokens / requirement | 7,789 / 8,667 / 9,596 | 7,061 / 8,013 / 6,660 |
We report medians and the ranges above — no run is cherry-picked.
Honest caveats
- Two metrics moved slightly the wrong way, both within noise. The confusion ratio (wasted ÷ total calls) rose +0.005 and the peak simultaneously-loaded tool count rose +2. Both are ≤ 0.12× the standard deviation of the leg-to-leg noise — the leg distributions almost entirely overlap, so there is no basis to read either as a regression. Every other metric moved in the expected direction.
- Completeness held. Both arms completed and playtested the full build in every run; the skill lowered the cost of getting there rather than trading away scope.
- This is a single game, single model, n=3. It is an indicative snapshot, not a guarantee. Treat the direction as the durable claim and the magnitude as specific to this build.
A different kind of efficiency number
This page measures a behavioral result — how a real agent’s run got cheaper with the skill installed — which depends on the game and the model, so it is a dated, re-measured snapshot. That is different from the toolkit’s structural cost. For the deterministic, exactly-reproducible number — how many context-window tokens the tool catalogue itself consumes — see token efficiency. The two are complementary: one is what the surface costs to register, the other is what a build costs to run.